From Arguments in Natural Language to Formalised Argumentation: Components, Prospects, and Problems

1
From Arguments in Natural Language to Formalised
ArgumentationComponents, Prospects, and
ProblemsICAIL Workshop on Natural Language
Engineering of Legal ArgumentationBarcelona,
SpainJune 8, 2009

• Adam Wyner Tom van Engers
• University College London University of Amsterdam

2
Outline

• The problem
• Current approaches
• Decomposing the problem
• Steps to building a tool

3
Example 1

• Informal public consultation debate concerning
  waste management
• H0 Householders should be charged for the
  amount of garbage they throw away.
• H1 This would provide a financial incentive to
  recycle more and waste less.
• H2 Families on benefits do not recycle and
  would not have any incentive to recycle. Why
  should I pay for the waste by other people?
• H3 Waste is created in other ways such
  supermarket packaging.
• H4 How about charging the food producers and
  manufacturers for waste, then reimburse people to
  recycle?
• H5 It would be better for the earth if we
  recycled. Besides, we are running out of
  available garbage dumps.
• H6 A lot of recycled stuff gets sent to the
  developing world, where it just poisons their
  environment. Look at China!
• H7 I already pay a lot for what I buy. Why
  should I pay more? All this ecological doom and
  gloom is just scare tactics to get more money out
  of us.

4
Example 2

• The employer is liable.
• The employer breached a duty of care.
• The employee had a work-related injury.
• The employer did not give the employee the safety
  instructions.
• A colleague says the employer did not give the
  employee the safety instructions.
• The employer is not liable.
• The employee was careless.
• The employee had no work-related injury.
• The employees injury was caused by poor physical
  condition.
• The colleague is not credible.
• The colleague is a friend of the employee.
• The employee was not careless.

5
An Argument is....

• A set of premises offered in support of a claim.
  Every human is mortal. Jill is a human.
  Therefore, Jill is mortal.
• Statements glued together so that the claims of
  one argument are the premises of another
  argument. A tree of statements. All the
  supporting statements for recycling.
• A debate for and against a particular claim. All
  the statements about recycling, pro and con.

6
Representations and Formalisations

• Argument schemes (Walton)
• Argumentation frameworks (Dung)
• Argument graphs of cases (Bench-Capon, Prakken)
• Propositional Logic and First-order Logic
• Structured opinion gathering (Atkinson)
• What we want to do
• Using linguistic input/output, to dynamically
  support the formal construction of arguments and
  argument interaction.

7
Problems

• Translating into a graphic or formalism is labour
  intensive.
• NL arguments are open to (mis)interpretation and
  misrepresentation.
• Arguments in unstructured text are linear in
  format, but conceptually non-linear.
• The parts of an argument can be in different
  orders, spread out in a text, and added over
  time.
• Hidden assumptions (enthymemes).
• Irrelevant claims can be made.
• Arguments are dialectical and defeasible.
  Subsequent statements attack parts of previous
  arguments (premises, claims, or rule
  applications).

8
Current Translation Method

• Argument AF
• in NL

A black box. Open the lid...
9
Current Translation Method
Argument AF in NL
Henry Prakken and Trevor Bench-Capon How do they
do it?
10
Analogy to FOL Translation

• How does one translate an expression in NL to
  FOL?
• Examples, patterns, and analogies. Taught in
  logic textbooks.
• Systematic translation and manipulation rules.
  Taught in formal semantics of natural language
  textbooks.

11
Goal Translation Method

• Ideal -- an implemented formal syntax and
  semantics of argumentation that translates from
  natural language input to a formal system reasons
  and back again. It would also fill in missing
  elements, ask questions, draw inferences....
• Subideal systems that provide some level of
  support for structured input in natural language
  and interface with a formal system. A series of
  (compromised) systems that build towards the
  ideal....
• Progress from the subideal to the ideal would
  give us greater understanding of argumentation.

12
Isnt NL the Problem?

• Natural language sentences are
• ambiguous (Jill saw a man with binoculars).
• context dependant (Jill went to the bank).
• have a complex syntax.
• have a rich lexicon.
• are vague, use metaphor, have anaphora...
• cannot be fully formalised in a logic.
• The problems are greater for multi-sentential
  structures and discourse/dialogue.

13
NL isnt the Problem!

• NL is the object of study. There are a range of
  subtopics to examine.
• Linguistics/Computational Linguistics where
  multi-sentential structures are translated into a
  logic.
• Every human is mortal. Jill is a human.
  Therefore, Jill is mortal.
• From human(jill), Forall x human(x) -gt
  mortal(x), infer mortal(jill).
• This is not machine learning/statistical NLP, but
  a formal syntax and semantics of natural
  language.
• Extend this approach to argumentation.

14
What Is Needed?

• Construction of an argument corpus.
  (Debateopedia)
• Annotated gold standards of arguments.
• Typical argument connectives and patterns
  (Argument Schemes, Pragma-dialectics and
  Rhetorical Argument Structure).
• Apply text mining and machine learning to
  identifying argument structures (and problems).
  A means to an end. (Moens and Mochales-Palau
  2009).
• Textual inference. (Dagan, Glickman, and Magnini
  2006)
• Identification of inconsistency. (de Marneffe,
  Rafferty, and Manning 2008)

15
What is Needed?

• Ontology of arguments (Rawhan).
• Parts of arguments (premises, exceptions, claims,
  and a rule).
• Relationships among the parts.
• Relationships between arguments. When does one
  argument subsume another?
• What is our knowledge of argument such that we
  can recognise that an argument is syntactically
  well-formed and that an inference semantically
  follows?
• Every human is mortal. Jill is a human.
  Therefore, Jill is mortal.
• Every human is mortal. The sun rises in the
  east. Therefore, the stock market is rigged.
• What is the role of lexical coherence?

16
What is Needed?

• Find inconsistent statements among previous
  statements.
• Incremental, defeasible argument construction.
• Find and eliminate redundancy over time.
• Find implications among statements over time.
• Construct complex discourses.
• Introduce presuppositions and background
  knowledge.
• Introduce argument relations (premise, claim,
  rule, contradiction....) to support argument
  input.
• Work in manageable subdomains of discourse. Not
  a general all purpose tool.

17
Argument Knowledge

• Is argument knowledge a catalogue of patterns,
  list of well-formed sentences in fixed
  structures, or is there an underlying abstract
  grammar? The latter would allow us to create
  novel grammatical arguments.

18
Test the Current View

• Some claim current formal analyses are adequate
  for argumentation.
• Examine this by providing means to translate from
  NL to AF.
• Make systematic translation part of the
  evaluation criteria for an adequate theory of
  argumentation?
• Example translation from NL expressions with
  quantifiers to FOL revealed the need for a
  different logic generalized quantifiers.

19
What is Useful?

• Controlled Languages (Right-sized/Appropriate-size
  d NL)
• Finite vocabulary (large).
• Limited grammatical structures (expressive).
• Predictive editor (support input of correct
  forms).
• Parse and translate to a logic (Discourse
  Representation Theory FOL).
• Web compatible representations.
• Attempto Controlled English
• http//attempto.ifi.uzh.ch/

20
ACE
21
Suggestions

• Look at work in Linguistics/Computational
  Linguistics from the last 5-10 years (textbooks,
  handbooks, ....) to understand current work on
  NL.
• Collaborate with a linguist/computational
  linguist.
• Pick a piece of the overall problem and work on
  it.
• Think of legal argument engineering as another AI
  problem autonomous vehicles, machine vision,
  walking robots, soccer playing robots, computer
  chess....
• Take an incremental approach.
• The whole problem may not be resolved anytime
  soon, but well learn some interesting things in
  the meantime as well as make some discoveries.
• Do knowledge engineering development with
  argument experts (e.g. field methods applied to
  argument natives).